[x265] [PATCH] [REVIEW PATCH]sa8dCost threshold to reduce number of modes in intra analysis
Steve Borho
steve at borho.org
Sat Jan 17 04:47:43 CET 2015
On 01/16, ashok at multicorewareinc.com wrote:
> # HG changeset patch
> # User Ashok Kumar Mishra<ashok at multicorewareinc.com>
> # Date 1421415068 -19800
> # Fri Jan 16 19:01:08 2015 +0530
> # Node ID 4c200eea0b2210fabb718d7ed8980c1af530ea3a
> # Parent f9a54da187139a7331f8048365673787cbf14711
> [REVIEW PATCH]sa8dCost threshold to reduce number of modes in intra analysis
>
> diff -r f9a54da18713 -r 4c200eea0b22 source/encoder/analysis.cpp
> --- a/source/encoder/analysis.cpp Thu Jan 15 11:49:26 2015 +0530
> +++ b/source/encoder/analysis.cpp Fri Jan 16 19:01:08 2015 +0530
> @@ -1187,13 +1187,13 @@
> (!m_param->bEnableCbfFastMode || md.bestMode->cu.getQtRootCbf(0)))
> {
> md.pred[PRED_INTRA].cu.initSubCU(parentCTU, cuGeom);
> - checkIntra(md.pred[PRED_INTRA], cuGeom, SIZE_2Nx2N, NULL);
> + checkIntra(md.pred[PRED_INTRA], cuGeom, SIZE_2Nx2N, NULL, md.bestMode[PRED_INTRA].sa8dCost * 3);
md.bestMode is a pointer to a single Mode in the md.pred[] array. We shouldn't be dereferencing
bestMode as an array.
> checkBestMode(md.pred[PRED_INTRA], depth);
>
> if (depth == g_maxCUDepth && cuGeom.log2CUSize > m_slice->m_sps->quadtreeTULog2MinSize)
> {
> md.pred[PRED_INTRA_NxN].cu.initSubCU(parentCTU, cuGeom);
> - checkIntra(md.pred[PRED_INTRA_NxN], cuGeom, SIZE_NxN, NULL);
> + checkIntra(md.pred[PRED_INTRA_NxN], cuGeom, SIZE_NxN, NULL, md.bestMode->sa8dCost);
> checkBestMode(md.pred[PRED_INTRA_NxN], depth);
> }
> }
> @@ -1549,6 +1549,12 @@
> }
> if (predInterSearch(interMode, cuGeom, bMergeOnly, true))
> {
> + const Yuv& fencYuv = *interMode.fencYuv;
> + Yuv& predYuv = interMode.predYuv;
> + int part = partitionFromLog2Size(cuGeom.log2CUSize);
> + interMode.distortion = primitives.cu[part].sa8d(fencYuv.m_buf[0], fencYuv.m_size, predYuv.m_buf[0], predYuv.m_size);
> + interMode.sa8dCost = m_rdCost.calcRdSADCost(interMode.distortion, interMode.sa8dBits);
> +
> /* predInterSearch sets interMode.sa8dBits, but this is ignored */
> encodeResAndCalcRdInterCU(interMode, cuGeom);
>
> diff -r f9a54da18713 -r 4c200eea0b22 source/encoder/search.cpp
> --- a/source/encoder/search.cpp Thu Jan 15 11:49:26 2015 +0530
> +++ b/source/encoder/search.cpp Fri Jan 16 19:01:08 2015 +0530
> @@ -1137,7 +1137,7 @@
> }
> }
>
> -void Search::checkIntra(Mode& intraMode, const CUGeom& cuGeom, PartSize partSize, uint8_t* sharedModes)
> +void Search::checkIntra(Mode& intraMode, const CUGeom& cuGeom, PartSize partSize, uint8_t* sharedModes, uint64_t sa8dThreshold)
> {
> CUData& cu = intraMode.cu;
>
> @@ -1148,7 +1148,13 @@
> cu.getIntraTUQtDepthRange(tuDepthRange, 0);
>
> intraMode.initCosts();
> - intraMode.distortion += estIntraPredQT(intraMode, cuGeom, tuDepthRange, sharedModes);
> + intraMode.distortion = estIntraPredQT(intraMode, cuGeom, tuDepthRange, sharedModes, sa8dThreshold);
> +
> + if (intraMode.distortion == MAX_INT)
> + {
> + intraMode.rdCost = MAX_INT64;
> + return;
> + }
> intraMode.distortion += estIntraPredChromaQT(intraMode, cuGeom);
>
> m_entropyCoder.resetBits();
> @@ -1396,7 +1402,7 @@
> updateModeCost(intraMode);
> }
>
> -uint32_t Search::estIntraPredQT(Mode &intraMode, const CUGeom& cuGeom, const uint32_t depthRange[2], uint8_t* sharedModes)
> +uint32_t Search::estIntraPredQT(Mode &intraMode, const CUGeom& cuGeom, const uint32_t depthRange[2], uint8_t* sharedModes, uint64_t sa8dThreshold)
> {
> CUData& cu = intraMode.cu;
> Yuv* reconYuv = &intraMode.reconYuv;
> @@ -1539,7 +1545,7 @@
>
> uint64_t paddedBcost = bcost + (bcost >> 3); // 1.12%
> for (int mode = 0; mode < 35; mode++)
> - if (modeCosts[mode] < paddedBcost || (mpms & ((uint64_t)1 << mode)))
> + if ((modeCosts[mode] < paddedBcost || (mpms & ((uint64_t)1 << mode))) && modeCosts[mode] < sa8dThreshold)
Perhaps we should be using:
paddedBcost = X265_MIN(paddedBcost, sa8dThreshold);
Then we wouldn't need an extra conditional in this loop, but it would
check all most probable modes unconditionally. It'll have to be
measured.
> updateCandList(mode, modeCosts[mode], maxCandCount, rdModeList, candCostList);
>
> /* measure best candidates using simple RDO (no TU splits) */
> @@ -1556,8 +1562,11 @@
> codeIntraLumaTSkip(intraMode, cuGeom, initTuDepth, absPartIdx, icosts);
> else
> codeIntraLumaQT(intraMode, cuGeom, initTuDepth, absPartIdx, false, icosts, depthRange);
> - COPY2_IF_LT(bcost, icosts.rdcost, bmode, rdModeList[i]);
> + COPY3_IF_LT(bcost, icosts.rdcost, bmode, rdModeList[i], intraMode.sa8dCost, candCostList[i]);
> }
> +
> + if (bcost == MAX_INT64)
> + return MAX_INT;
> }
>
> /* remeasure best mode, allowing TU splits */
> diff -r f9a54da18713 -r 4c200eea0b22 source/encoder/search.h
> --- a/source/encoder/search.h Thu Jan 15 11:49:26 2015 +0530
> +++ b/source/encoder/search.h Fri Jan 16 19:01:08 2015 +0530
> @@ -163,7 +163,7 @@
> void invalidateContexts(int fromDepth);
>
> // full RD search of intra modes. if sharedModes is not NULL, it directly uses them
> - void checkIntra(Mode& intraMode, const CUGeom& cuGeom, PartSize partSize, uint8_t* sharedModes);
> + void checkIntra(Mode& intraMode, const CUGeom& cuGeom, PartSize partSize, uint8_t* sharedModes, uint64_t sa8dThreshold = MAX_INT);
it shouldn't matter, no 32x32 sa8d is going to return a value larger
than MAX_INT, but I suppose this should be MAX_INT64 given the variable
is uint64_t.
> // select best intra mode using only sa8d costs, cannot measure NxN intra
> void checkIntraInInter(Mode& intraMode, const CUGeom& cuGeom);
> @@ -204,7 +204,7 @@
> void saveResidualQTData(CUData& cu, ShortYuv& resiYuv, uint32_t absPartIdx, uint32_t depth);
>
> // RDO search of luma intra modes; result is fully encoded luma. luma distortion is returned
> - uint32_t estIntraPredQT(Mode &intraMode, const CUGeom& cuGeom, const uint32_t depthRange[2], uint8_t* sharedModes);
> + uint32_t estIntraPredQT(Mode &intraMode, const CUGeom& cuGeom, const uint32_t depthRange[2], uint8_t* sharedModes, uint64_t sa8dThreshold);
>
> // RDO select best chroma mode from luma; result is fully encode chroma. chroma distortion is returned
> uint32_t estIntraPredChromaQT(Mode &intraMode, const CUGeom& cuGeom);
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
--
Steve Borho
More information about the x265-devel
mailing list